AI agents hit Anthropic, GitHub’s agent surge, and Gemini 4 Carbon

Compact Conversations for 2026-10-28: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: AI agents hacked Anthropic through third-party software

Cybersecurity startup Cathedral used AI agents to break into Anthropic’s internal systems via third-party business software, accessing Slack messages, salary data, and more within hours. Anthropic confirmed limited access to sales, HR, and IT systems but disputes the depth and timeline, saying source code and model weights were never exposed.

Why it matters: Even a leading AI lab is vulnerable to AI-powered attacks through external software dependencies, highlighting a growing threat for all enterprises.

Source: The Decoder

Number to Know: GitHub Octoverse 2026: 220 million merged PRs per month

GitHub’s Octoverse 2026 report shows merged pull requests tripled to 220 million per month, commits more than tripled to 7.38 billion per month, and new repositories doubled to 30.3 million per month—driven by agentic engineering.

Why it matters: These numbers quantify the massive impact of AI coding agents on software development activity.

Source: Simon Willison’s Weblog

The Feed

Google unveils Gemini 4 Carbon, its strongest model with state-of-the-art cybersecurity

Gemini 4 Carbon tops Artificial Analysis’ Intelligence Index and claims state-of-the-art performance in vulnerability detection, exploitation, and patching. Access starts with a small group of cyber defenders this week, with broader availability later.

Why it matters: Google is positioning Carbon as a leader in AI-driven cyber defense, with pricing matching Argon at $2 per million input tokens and $10 per million output tokens.

Source: Google

Google opens Android development to third-party AI agents

Android Studio’s new ‘Bring Your Own Agent’ feature supports over 40 coding agents, including Claude Agent and Codex. Google says new terminal tools help agents finish UI tasks 77% faster while using up to three times fewer tokens.

Why it matters: Third-party AI agents are becoming first-class citizens in mobile development, expanding options for developers.

Source: The Decoder

AI maths advances raise fears over security of global encryption

The Financial Times reports that AI’s growing mathematical prowess is forcing a rethink of cybersecurity planning that has long focused on the quantum threat.

Why it matters: AI’s rapid math advances present a second, faster-moving risk to global encryption, alongside quantum computing.

Source: Financial Times

GPT-6.1 Terra matches OpenAI’s heavyweight at a tenth of the price

OpenAI’s mid-tier GPT-6.1 Terra matches the flagship GPT-6 Astra on agentic benchmarks at roughly a tenth of the token price, though it’s weaker on hard coding.

Why it matters: The gap between mid-tier and flagship models is collapsing into a price difference, making high-end performance more accessible.

Source: beehiiv.com

One Thing to Try

From a popular Reddit thread: put every prompt in git with a version tag, build a small eval set of 20-50 real inputs with expected outputs, and review changes like pull requests—diffing eval results instead of just prompt text. Log outputs to catch drift and roll back easily. LLM-as-judge works for format but gets shaky on nuance.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead: cybersecurity startup Cathedral used AI agents to break into Anthropic’s internal systems through third-party business software, as part of Anthropic’s own bug bounty program. Cathedral, founded by former employees of the government efficiency agency DOGE, says its agents found Slack messages, salary data, meeting transcripts, and job postings in under twelve hours. Host B: [thoughtful] Anthropic confirmed the researchers reached a limited set of sales, HR, and IT systems, including some customer and partner contact info. But it disputes the timeline and the depth, saying the probing stretched over several days, and that Claude chats, source code, and model weights were never exposed. Anthropic shut down the affected access points the same day it found the breach and notified federal authorities. Host A: [curious] The two sides are also sparring over the rules. Anthropic says Cathedral violated the bug bounty’s safe harbor by targeting third-party providers and pulling data. A former OpenAI employee claimed publicly that Anthropic is downplaying it, saying Cathedral got full admin access to Slack. Anthropic’s red team lead pushed back on that, though he wrote that the industry now has to reinvent its security every year.

Host B: One number to know today: 220 million pull requests merged every month, per GitHub’s new Octoverse 2026 report, announced at GitHub Universe. That’s triple the rate from less than a year ago. [with emphasis] Monthly commits more than tripled too, to 7.38 billion, with most of that jump in the second half of the year. And new repositories doubled to 30.3 million a month. Host A: [curious] GitHub frames that as the impact of agentic engineering on software development, and the curves do spike right along with it.

Host A: [quick aside] Google launched a new model focused on cyberdefense. Gemini 4 Carbon is the company’s newest frontier model, two weeks after Gemini 4 Argon, and it tops Artificial Analysis’ latest Intelligence Index. Google says it sets state-of-the-art results on benchmarks for finding vulnerabilities, validating them through exploitation, and patching at the root cause. Access starts this week with a small subset of cyber defenders through its Fairwind Program, before broader developer availability. Pricing matches Argon: two dollars per million input tokens, ten for output. Host B: In other agent news, Google is opening Android development to third-party AI agents. Android Studio’s new Bring Your Own Agent feature supports more than forty coding agents, including Anthropic’s Claude Agent and OpenAI’s Codex. [with a small lift] Google also says new terminal tools help agents finish UI tasks 77 percent faster while using up to three times fewer tokens. Host A: The Financial Times reports that AI’s growing mathematical prowess is raising fears over the security of global encryption, forcing a rethink of cyber security planning that has long focused on the quantum threat. Host B: [lighter] And one more model story, from AI Breakfast: OpenAI’s GPT-6.1 Terra, a mid-tier model for visual reasoning and computer use, matches the flagship GPT-6 Astra on its agentic benchmark at roughly a tenth of the token price. [skeptical] It’s weaker on hard coding, and OpenAI still points the toughest work at its bigger models.

Host A: For today’s One Thing to Try, a prompt-engineering habit from a popular Reddit thread: treat your prompts like production code. [conversational] The poster took Karpathy’s idea that prompts are programs literally, and put every prompt in git with a version tag. Then they built a small eval set of twenty to fifty real inputs with expected outputs that runs before any change ships. Host B: [thoughtful] The part that stands out is reviewing prompt changes like pull requests, but diffing the eval results instead of just the prompt text. They also log outputs to catch drift when a provider updates a model, and can roll back to the last good version in one step. The honest caveat: LLM-as-judge works fine for format checks but gets shaky on nuance. Pick one prompt your team actually relies on and give it a version history and a tiny eval set.

Host A: That’s Compact Conversations for Wednesday. More AI news tomorrow. Until then, happy prompting.